Debian Patches

Status for hipblas/7.2.4-1~exp1

Patch Description Author Forwarded Bugs Origin Last update
0001-remove-immintrin-header.patch remove immintrin header
This header is unused, but it breaks the build on arm64 and ppc64el
because it is not available on those platforms.		 Cordell Bloor <cgmb@slerp.xyz> yes 2024-03-22
0002-Disable-omp.patch Disable omp Christian Bayle <bayle@debian.org> not-needed 2026-06-11
0003-use-generic-blas-and-lapack.patch use generic blas and lapack
The upstream project typically uses either the AOCL BLIS library or the
Netlib BLAS library as the reference implementation in the test suite on
Linux. However, the OpenBLAS library is used by upstream on Windows. It
would be nice to use OpenBLAS on Debian for performance reasons (as the
test suite is heavily CPU-bound), however, the Netlib implementation
seems to be more reliable for achieving a full suite of passing tests.		 Cordell Bloor <cgmb@debian.org> no 2026-06-11
0004-Use-local-mathjax.patch Use local mathjax Christian Bayle <bayle@debian.org> not-needed 2025-10-03
0005-Gtest-add-verbose-flag-to-prevent-timeout.patch Add verbose flag to prevent build timeout on Launchpad Talha Can Havadar <talha.can.havadar@canonical.com> no 2026-06-11
0006-drop-f16c-instructions.patch drop f16c instructions
These instructions are x86-specific. The __half implementation from the HIP
runtime can be used instead.		 Cordell Bloor <cgmb@slerp.xyz> no 2026-06-11
0007-optional-git.patch optional git
Git is only used to print a message during the build. The message is
not useful when building from a tarball or from the salsa git repo.		 Cordell Bloor <cgmb@debian.org> no 2025-11-18
0008-fix-ambiguous-__half-constructor.patch fix ambiguous __half constructor
This mostly fixes errors like this one

64 | const Ta h_alpha(1), h_zero(0);
| ^~~~~~~

which occurs when building using __half as the floating point type.		 Cordell Bloor <cgmb@debian.org> no 2025-11-18
0009-enable-changing-test-data-dir.patch enable changing test data dir
This patch is based off of and almost identical to the
'Enable-changing-directory-for-test-data.patch' in rocblas, written by
Christian Kastner.

It allows for test data to be installed at an arbitrary directory
defined by INSTALL_TEST_DATA_DIR so that the platform independent data
can be packaged separately from the platform-dependant test binaries.		 Tim Flink <Tim.Flink@amd.com> not-needed 2026-06-11
0010-remove-readme-from-doxygen-sources.patch remove readme.md as doxygen source
Doxygen issues warnings when it uses the README.md file as input.

This removes README.md as a source for Doxygen to keep that warning
from causing the doc generation process to abort.		 Tim Flink <Tim.Flink@amd.com> no 2026-01-28
0011-fix-arm64-immintrin-include.patch Guard immintrin.h include for x86 architectures only The header <immintrin.h> provides x86 SIMD intrinsics (SSE, AVX, etc.) and
is only available on x86/x86_64 platforms. Including it unconditionally
causes build failures on arm64 with the error:
"This header is only meant to be used on x86 and x64 architecture"
.
Since hipblaslt_math.hpp does not actually use any x86 intrinsics, the
include appears to be unused. This patch guards it with architecture
detection macros to allow building on arm64 while preserving the include
on x86 systems where it might be needed.

===================================================================		 Igor Luppi <igor.luppi@canonical.com> invalid 2026-02-06
0012-fix-tests-host-vector-type-and-lapack-bits-size.patch Fix failing batched tests on multiple test suites Certain test implementations were causing crashes and false failures
during autopkgtest runs.
.
The first issue was a likely typo in a `host_vector` type declaration.
It was meant to hold integer status codes (`info` values) but was
declared using the matrix data template `<T>`. This caused a type
mismatch during validation, which artificially accumulated an error
penalty on each test iteration.
.
The second issue was a memory misalignment caused by reading 32-bit
LAPACK outputs into 64-bit integer arrays, leading to out-of-bounds
writes and segmentation faults across multiple test suites. The fix
now checks if the linked LAPACK library uses 32-bit or 64-bit
integers and safely unpacks the data accordingly to prevent
buffer overflows.
.
Because these errors are caused by handling LAPACK's 32-bit/64-bit integer
ABIs, this patch was originally expected to be obsoleted in 7.2.0, which
announced a switch to `liblapack64-dev` (ILP64) for the test clients.
However, the upstream fix (commit 9252c306) did not make it into the
7.2.4 release tarball: `clients/include/solver/testing_gels_batched.hpp`
still declares `info_res` / `info` as `host_vector<T>` (the matrix data
template) instead of `host_vector<int>`, and the LAPACK 32/64-bit
detection logic in `testing_getri_batched.hpp` and
`testing_getrs_batched.hpp` is also still missing. The patch is
therefore still empirically required at 7.2.4. Re-evaluate at the next
upstream bump and drop once the upstream tree carries the fix.

===================================================================		 Bruno Bernardo de Moura <bruno.moura@canonical.com> not-needed upstream, https://github.com/ROCm/rocm-libraries/commit/9252c3067ee2afc565c7064e3d3771ed7d4a4b6b 2026-06-15

All known versions for source package 'hipblas'

Links